Tutorials

Downloads:

• slides
• code

Presenter: Dr. Lilach Goren Huber, ZHAW School of Engineering

Description: Developing deep learning algorithms for predictive maintenance of industrial systems is a growing trend in numerous application fields. Whereas applied research methods have been rapidly advancing, implementations in commercial systems are still lagging behind. One of the main reasons for this delay is the fact that most methodological advances have been focusing on the development of data-driven algorithms for fault detection, diagnosis, or prognosis, ignoring some of the crucial aspects that are required for scaling these algorithms to large fleets of multi-component heterogeneous machines under varying operating conditions, and making sure that their implementation is technically feasible.

In this tutorial, we will elaborate on some of these aspects and discuss possible approaches to address them. We will provide the background to data analytical techniques that enable the scalable deployment of deep learning algorithms in commercial machine fleets. Some examples are transfer learning, fleet-level algorithms, physics-informed deep learning, and uncertainty quantification. We will demonstrate these general concepts using concrete use-cases that apply them to operational data from commercial machine fleets.

Speaker Bio: Lilach Goren Huber is a senior scientist and project leader at the school of engineering of the Zurich University for Applied Sciences in Switzerland since 2013. After completing her PhD in physics at the Weizmann Institute of Science in Israel she joined an Israeli start-up as a physicist and data scientist, developing algorithms for remote sensing of wind speed and direction. At the Zurich University for Applied Sciences, she focuses on machine learning and deep learning algorithms for intelligent operation and maintenance of machines and infrastructures, leading research and development projects in collaboration with industry partners and the public sector in diverse application fields.

Downloads:

• • slides (PDF; no animations)
  • slides (ZIP’d PPTX; with animations)

Presenter: Neil Eklund, Novity

Description: Binary classifiers are incredibly common in asset health management: “normal or abnormal”, or “faulty or no-fault” are core questions in PHM. It is very common in the PHM literature to illustrate performance with the Receiver Operating Characteristics (ROC) curve, and compare models using the area under the ROC curve, or “AUC” (“area under curve”) statistic. However, the vast majority of PHM data sets are wildly imbalanced, e.g., 29,000 normal operations, and 13 faults. ROC plots can be misleading for the type of imbalanced data found in asset health management, due to an intuitive but incorrect interpretation of false alarm rate. This tutorial will provide an overview of statistical and graphical methods for comparing classifiers, and will highlight methods that provide clear visual cues and an accurate and intuitive interpretation of practical classifier performance using both real-world examples and simulation data.

Speaker Bio: Neil Eklund is one of the most innovative, forward-thinking researchers in the field of Prognostics and Health Management (PHM). With 20 years of deep technical experience in machine learning, data science, and industrial analytics, Neil serves as Principal Data Scientist at Novity, the industrial predictive maintenance venture recently launched by Xerox PARC. Before joining Novity, Dr. Eklund served as the Chief Data Scientist and Scientific Advisor at Schlumberger, where his work led to the creation of a new PHM-focused business segment, Technology Lifecycle Management. Prior to that, he spent 14 years in the Machine Learning Laboratory at GE Global Research where he was responsible for bringing advanced analytics as a shared service to GE businesses for better and faster utilization of the company’s industrial data assets. Dr. Eklund holds a Ph.D. in Machine Learning and he is also a PHM Society Fellow.

Presenter: Professor Sankaran Mahadevan, Vanderbilt University

Description: The digital twin paradigm integrates information obtained from sensor data, physics models, as well as operational and inspection/maintenance/repair history of a physical system or component of interest. As more and more data becomes available, the resulting updated model becomes increasingly accurate in predicting future behavior of the system, and can potentially be used to support several objectives, such as sustainment, mission planning, and operational maneuvers. This tutorial will present recent advances in digital twin methodologies to support all three objectives, based on several types of computation: current state diagnosis, model updating, future state prognosis, and decision-making. All these computations are affected by uncertainty regarding system properties, operational parameters, usage and environment, as well as uncertainties in data and the prediction models. Therefore, the tutorial will also address uncertainty quantification in diagnosis and prognosis (considering both aleatory and epistemic uncertainty sources), and decision-making under uncertainty. Scaling up the probabilistic digital twin methodology to support real-time decision-making is a challenge, and several strategies that combine recent advances in sensing, computing, data fusion and machine learning to enable the scale-up will be discussed. Several use cases related to aircraft, rotorcraft, marine vessels, and additive manufacturing will be presented.

Speaker Bio: Professor Sankaran Mahadevan (Vanderbilt University, Nashville, TN, USA) has more than thirty years of research and teaching experience in uncertainty quantification, risk and reliability analysis, machine learning, structural health diagnosis and prognosis, and decision-making under uncertainty. He has applied these methods to a variety of structures, materials and systems in civil, mechanical and aerospace engineering. His research has been extensively funded by NSF, NASA, DOE, DOD, FAA, NIST, as well as GM, Chrysler, GE, Union Pacific, and Mitsubishi, and he has co-authored two textbooks and 330 peer-reviewed journal papers. During the past decade, he has been at the forefront of academic research on digital twin methodologies for aircraft, rotorcraft, ship structures, and additive manufacturing, funded by FAA, U.S. Air Force, U. S. Army, and NIST. Professor Mahadevan is currently Managing Editor of the ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, and has served as General Chair of several prominent conferences such as the AIAA SDM Conference, AIAA Non-Deterministic Approaches Conference, ASCE Engineering Mechanics Conference, and three PHM Society Annual Conferences, including PHM 2022. He is a Fellow of AIAA, Engineering Mechanics Institute (ASCE), and PHM Society.